Neural radiance fields (NeRF) have established a new paradigm for 3D scene reconstruction, with subsequent work achieving high-quality real-time rendering. However, reconstructing large-scale scenes from oblique aerial photography presents unique challenges, such as varying spatial scale distributions and a constrained range of tilt angles, often resulting in high memory consumption and reduced rendering quality at extrapolated viewpoints. To address these issues, we propose a novel approach named Oblique-MERF to accommodate the distinctive characteristics of oblique photography datasets and support real-time rendering on various common devices. Firstly, an innovative adaptive occupancy plane is proposed to constrain the sampling space. Additionally, we propose a smoothness regularization loss for view-dependent color to enhance the MLP's ability to generalize to untrained viewpoints. Experimental results demonstrate that Oblique-MERF reduces VRAM usage by approximately 40% while maintaining competitive rendering quality compared to baseline methods, and achieves higher frame rates with more realistic rendering even at untrained extrapolated viewpoints. Project page: https://ustc3dv.github.ioIOblique-MERFI

Optical Coherence Tomography (OCT) is a high-resolution 3D imaging technique widely used in medical diagnostics. However, its increasing resolution leads to massive data volumes, creating challenges in storage, transmission, and real-time access—issues that are further intensified in Internet of Things (IoT) environment. In such settings, where portable OCT devices operate at the edge in remote or resource-limited environments, efficient and secure data handling is critical for real-time diagnostics, remote consultations, and cloud integration. Existing compression methods often struggle to balance efficiency, image quality, and the preservation of diagnostic features, limiting their applicability in constrained IoT infrastructures. To address these challenges, we propose OCT-CSNet, a deep compressed sensing (CS) model tailored for OCT images. By integrating deep learning with traditional CS theory, our model leverages non-parametric prior features to guide both compression and reconstruction stages. In the compression stage, the dual-path compression module efficiently captures the inherent sparsity and structural coherence of OCT images while preserving critical differential information. In the reconstruction stage, we fuse the Proximal Gradient Descent (PGD) algorithm with a deep neural network, incorporating data-driven priors into the iterative optimization process. Extensive experiments show that OCT-CSNet can improve the reconstruction quality under different compression ratios while retaining semantic features that are critical to downstream tasks. In addition, its high computational efficiency and low memory footprint also provide potential for its application in IoT-driven imaging systems.

3D editing, particularly involving realistic color editing, plays a crucial role in various multimedia domains, such as augmented reality and filmmaking. Traditional 3D reconstruction methods encounter challenges in achieving high-fidelity reconstruction for complex scenes. In recent years, methods based on implicit 3D representations, like Neural Radiance Fields (NeRF), have demonstrated effectiveness in rendering complex scenes. However, these methods face difficulties in interactively editing scene colors and often exhibit slow processing speeds. Addressing these challenges, we propose the PaletteGaussian framework for interactive color editing and real-time rendering based on a palette and 3D Gaussian Splatting (3DGS). First, we introduce a two-stage training strategy to ensure rendering quality and enhance the accuracy of object extraction in the scene. Next, we present an image-driven learning-based approach, I-learning, for convenient interactive color editing driven by both images and text. Finally, we perform parameter baking to achieve real-time rendering. In summary, PaletteGaussian supports two editing levels, scene-level and object-level, offering three interaction modes: manual, image-driven, and text-driven editing. It enables high-resolution real-time rendering. Our comprehensive experiments demonstrate that PaletteGaussian exhibits efficient performance, diverse interaction modes, and realistic color editing.

To identify similar diseases has significant implications for revealing the etiology and pathogenesis of diseases and further research in the domain of biomedicine. Currently, most methods for the measurement of disease similarity utilize either associations of ontological disease concepts or functional interactions between disease-related genes. These methods are heavily dependent on the ontology, which are not always available, and the selection of datasets. Moreover, many methods suffer from a drawback that they only use a single metric to evaluate disease similarity from an individual data source, which may result in biased conclusions without consideration of other aspects. In this study, we proposed a novel ontology-independent framework, namely RADAR, for learning representations for diseases to deduce their similarities from an integrative perspective. By leveraging the associations between diseases and disease-related biomedical entities, a disease similarity network was built under various metrics. Then, a multi-layer disease similarity network was constructed by integrating multiple disease similarity networks derived from multiple data sources, where the representation learning was derived to provide a comprehensive evaluation of disease similarities. The performance of RADAR was assessed by a benchmark disease set and 100 random disease sets. Experimental results demonstrated that RADAR can detect similar diseases effectively.

Background: Machine Learning (ML) has been widely used as a powerful tool to support Software Engineering (SE). The fundamental assumptions of data characteristics required for specific ML methods have to be carefully considered prior to their applications in SE. Within the context of Continuous Integration (CI) and Continuous Deployment (CD) practices, there are two vital characteristics of data prone to be violated in SE research. First, the logs generated during CI/CD for training are imbalanced data, which is contrary to the principles of common balanced classifiers; second, these logs are also time-series data, which violates the assumption of cross-validation. Objective: We aim to systematically study the two data characteristics and further provide a comprehensive evaluation for predictive CI/CD with the data from real projects. Method: We conduct an experimental study that evaluates 67 CI/CD predictive models using both cross-validation and time-series-validation. Results: Our evaluation shows that cross-validation makes the evaluation of the models optimistic in most cases, there are a few counter-examples as well. The performance of the top 10 imbalanced models are better than the balanced models in the predictions of failed builds, even for balanced data. The degree of data imbalance has a negative impact on prediction performance. Conclusion: In research and practice, the assumptions of the various ML methods should be seriously considered for the validity of research. Even if it is used to compare the relative performance of models, cross-validation may not be applicable to the problems with time-series features. The research community need to revisit the evaluation results reported in some existing research.